Method and system for identifying corrosion defect of power transformation fitting

Abstract

The invention discloses a power transformation fitting corrosion defect identification method and system, and belongs to the technical field of electric power. According to the method, the image data is adjusted through the environment data to obtain the standardized image data, and interference of environment factors such as illumination change, shielding and angle deviation in a complex field on the image data can be eliminated; then visual features are extracted from the standardized image data, the target type and the installation position of the power transformation fitting are determined according to the visual features, and the structural features and the installation scenes of different power transformation fittings can be accurately matched; then, an environment corrosion associated knowledge group is determined in combination with the power transformation fitting knowledge graph, and then corrosion associated knowledge is determined, so that fitting field knowledge and field data are deeply fused, and the risk of missing detection caused by incomplete information is reduced; and then defect identification is carried out based on the electric power defect identification graph model to obtain a corrosion defect identification result, so that accurate identification of the corrosion defect of the power transformation fitting is realized, and the accuracy of identification of the corrosion defect of the power transformation fitting is improved.

Description

Technical Field

[0001] This invention belongs to the field of power technology, and in particular relates to a method and system for identifying corrosion defects in substation fittings. Background Technology

[0002] In modern power transmission systems, substation fittings, as key components connecting, fixing, and protecting transmission lines, directly determine the operational safety and long-term reliability of the entire power grid. These fittings are typically exposed to complex and variable outdoor environments for extended periods, facing the combined effects of high temperatures, high humidity, salt spray, industrial dust, and ultraviolet radiation. Under these harsh conditions, the surface of the fittings is highly susceptible to corrosion, significantly reducing their mechanical strength, deteriorating conductivity, increasing contact resistance, and subsequently triggering a chain reaction of problems such as localized overheating and arcing. In severe cases, corrosion can even cause fitting breakage, conductor detachment, or equipment short circuits, posing a significant threat to power grid stability and imposing substantial maintenance costs and safety risks on power companies. Therefore, achieving accurate and efficient identification of corrosion defects in substation fittings is one of the core aspects of ensuring the safe operation of the power grid.

[0003] Currently, the industry primarily relies on computer vision methods based on single-vision perception to identify corrosion defects in substation fittings. This involves acquiring on-site images of the fittings and then detecting and judging corrosion areas based on image color and texture features. However, the actual environment is complex and variable. Drastic fluctuations in light intensity, localized obstructions caused by the equipment itself or surrounding components, and inconsistencies in shooting distance and angle can all lead to noise, shadows, and distortion in the images, severely impacting image quality and feature consistency, ultimately weakening the accuracy and robustness of corrosion identification. Furthermore, existing methods largely depend solely on visual data and fail to effectively incorporate domain-specific knowledge of substation fittings. For example, the differences in structural design, material composition, and functional characteristics among different types of fittings, as well as the inherent correlation between their installation location and environmental stress, make it difficult to comprehensively capture subtle corrosion patterns and to make adaptive judgments based on fitting type and environmental background. This leads to missed or misjudged corrosion defects when facing diverse on-site scenarios. Summary of the Invention

[0004] The present invention aims to provide a method and system for identifying corrosion defects in substation fittings to solve the above-mentioned technical problems. By adjusting image data through environmental data and constructing environmental corrosion-related knowledge groups, the accuracy of identifying corrosion defects in substation fittings is improved.

[0005] To address the aforementioned technical problems, this invention provides a method for identifying corrosion defects in substation fittings, comprising: acquiring image data of the substation fittings and several types of environmental data, and adjusting the image data based on the environmental data to determine standardized image data of the substation fittings; extracting visual features of the substation fittings from the standardized image data, and determining the target type and installation location of the substation fittings based on the visual features; performing node retrieval on a preset substation fitting knowledge graph based on the target type, installation location, and several types of environmental data of the substation fittings to determine environmental corrosion-related knowledge groups of the substation fittings; constructing several corrosion influence paths of the substation fittings based on the environmental corrosion-related knowledge groups, and extracting several key corrosion knowledge nodes from the substation fitting knowledge graph of the corrosion influence paths, and then constructing corrosion-related knowledge of the substation fittings based on the key corrosion knowledge nodes; performing path recognition on a preset power defect identification graph model based on visual features and corrosion-related knowledge to determine the effective path groups of the substation fittings, and confirming the defects of the substation fittings by combining visual features and corrosion-related knowledge to obtain corrosion defect identification results.

[0006] This invention standardizes image data by adjusting it with environmental data, eliminating interference from environmental factors such as lighting changes, occlusion, and angular deviations in complex environments. Visual features are then extracted from the standardized image data to determine the target type and installation location of substation fittings. This accurately matches the structural characteristics and installation scenarios of different substation fittings, effectively eliminating identification biases caused by misidentification of fitting types or unclear installation locations. Based on the target type, installation location, and environmental data, and combined with a substation fitting knowledge graph, environmental corrosion-related knowledge groups are determined, and corrosion impact paths are constructed. This identifies key corrosion knowledge nodes related to corrosion impacts and establishes corrosion-related knowledge, deeply integrating fitting domain knowledge with field data. This allows for comprehensive capture of subtle corrosion patterns in substation fittings under different scenarios, reducing the risk of missed detections due to incomplete information. Finally, based on a power defect identification graph model, combined with visual features and corrosion-related knowledge, defect identification is performed to obtain corrosion defect identification results, achieving accurate identification of substation fitting corrosion defects and improving the accuracy of corrosion defect identification.

[0007] Accordingly, this invention provides a system for identifying corrosion defects in substation fittings, comprising: an image standardization module, a visual feature extraction module, a knowledge graph retrieval module, a corrosion association knowledge construction module, and a corrosion defect identification result acquisition module. The image standardization module acquires image data of the substation fittings and several types of environmental data, and adjusts the image data based on the environmental data to determine standardized image data of the substation fittings. The visual feature extraction module extracts visual features of the substation fittings from the standardized image data and determines the target type and installation location of the substation fittings based on the visual features. The knowledge graph retrieval module is used to determine the target type, installation location, and several environmental factors of the substation fittings. The system uses data to perform node retrieval on a pre-defined substation fitting knowledge graph to identify environmental corrosion-related knowledge groups for the substation fittings. A corrosion-related knowledge construction module is used to construct several corrosion impact paths for the substation fittings based on these environmental corrosion-related knowledge groups, and extract several key corrosion knowledge nodes from the substation fitting knowledge graph for these paths. Then, based on these key corrosion knowledge nodes, the system constructs corrosion-related knowledge for the substation fittings. A corrosion defect identification result acquisition module is used to perform path identification on a pre-defined power defect identification graph model based on visual features and corrosion-related knowledge, determine the effective path groups for the substation fittings, and combine visual features and corrosion-related knowledge to confirm defects in the substation fittings, obtaining corrosion defect identification results.

[0008] This system standardizes image data by adjusting it with environmental data, thus mitigating the interference of environmental factors. Visual features are then extracted from this standardized image data to determine the target type and installation location of substation fittings. This allows for precise matching of the structural characteristics and installation scenarios of different substation fittings, effectively eliminating identification biases caused by confusion in fitting types or unclear installation locations. Next, based on the target type, installation location, and environmental data of the substation fittings, and combined with a substation fitting knowledge graph, environmental corrosion-related knowledge groups are identified. This constructs corrosion impact paths, identifying key corrosion knowledge nodes related to corrosion impacts and determining corrosion-related knowledge. This deep integration of fitting domain knowledge with field data enables comprehensive capture of subtle corrosion patterns in substation fittings under different scenarios, reducing the risk of missed detections due to incomplete information. Finally, based on a power defect identification graph model, combined with visual features and corrosion-related knowledge, defect identification is performed to obtain corrosion defect identification results, improving the accuracy of substation fitting corrosion defect identification. Attached Figure Description

[0009] Figure 1 A flowchart illustrating the steps of a method for identifying corrosion defects in substation fittings provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of a substation fitting corrosion defect identification system provided in an embodiment of the present invention. Detailed Implementation

[0010] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0011] Example 1 Please refer to Figure 1 , Figure 1 The flowchart of a method for identifying corrosion defects in substation fittings provided in this embodiment of the invention includes steps S1 to S5.

[0012] Step S1: Acquire image data of the substation fittings and several types of environmental data, and adjust the image data based on the environmental data to determine the standardized image data of the substation fittings. This includes steps S101 to S108. Step S101: Acquire image data of the substation fittings and several types of environmental data, including: light intensity data, temperature data, humidity data, and dust concentration data. Step S102: Divide the image data into several image regions and calculate the influence factor of each type of environmental data in each image region. Step S103: Based on the environmental data and influence factors, calculate the correlation mapping values ​​between light intensity data and light intensity of each image region, temperature data and temperature of each image region, humidity data and humidity of each image region, and dust concentration data and dust concentration of each image region. Step S104: Based on the light intensity correlation mapping values ​​and combined with the preset reference brightness of each image region, calculate the light deviation value of each image region. Step S105: Obtain the blurriness of each image region and, combined with the humidity-related mapping value, determine the amount of humidity-induced blurriness reduction for each image region. Step S106: Obtain the contrast deviation value of each image region and, combined with the temperature-related mapping value, determine the contrast compensation value for each image region. Step S107: Extract the edge sharpness data of each image region, determine the sharpness reduction rate of each image region based on the edge sharpness data, and, combined with the dust concentration-related mapping value, determine the sharpness restoration coefficient for each image region. Step S108: Based on the illumination deviation value, contrast compensation value, humidity-induced blurriness reduction amount, and sharpness restoration coefficient of each image region, adjust the image data to determine the standardized image data for substation fittings.

[0013] In one optional embodiment, high-definition industrial cameras mounted on drones, fixed monitoring pan-tilt units, and inspection robots collect image data of substation fittings. These cameras cover a shooting angle range of 0° to 180° to ensure that key areas such as the front, sides, and back of the substation fittings are captured. Subsequently, sensor arrays are deployed in different areas of the substation. These arrays can monitor in real time data affecting the corrosion of substation fittings, such as light intensity, temperature, humidity, and dust concentration. Appropriate sampling frequencies are set during environmental data collection. After collecting environmental and image data, outlier removal is performed on the environmental data, using the 3σ criterion for judgment. Environmental data exceeding the normal range was cut off and removed. Furthermore, taking a 220kV substation as an example, a drone carrying a high-definition industrial camera captured images of the suspension clamp hardware within the substation from three angles: 30°, 90°, and 150°, acquiring three sets of image data with a resolution of 4096×3072. Simultaneously, a fixed monitoring pan-tilt unit and an inspection robot supplemented the images from 0° and 180° angles, acquiring a total of five sets of image data. The sensor array monitored the area at that time, finding the light intensity to be 8500 lux (strong light environment), humidity to be 65%RH, temperature to be 32℃, dust concentration to be 0.8 mg / m³, and salt spray concentration to be 0.05 mg / m³.

[0014] In one optional embodiment, the image data is divided into several image regions of 64×64 pixels (if the pixel density is high in critical parts of the substation fittings (such as bolt connections), it can be adaptively adjusted to a finer region of 32×32 pixels). Then, a spatial interpolation algorithm (Kriging interpolation method is used in this embodiment) is used to map the light intensity data, temperature data, humidity data, and dust concentration data to each image region. Normalization is then performed to obtain an influence factor for each environmental data type, with values ​​ranging from 0 to 1, in each image region. Finally, the image regions are used as row indices (e.g., the first row corresponds to the first image region, denoted as...). The Nth row corresponds to the Nth image region, denoted as... Using environmental data (light intensity L, temperature T, humidity H, dust concentration D) as column indices, an N×4 environmental association mapping matrix M is constructed, with matrix elements... Indicates the first The first image region The associated mapping value of each environmental data ( =1 indicates a light intensity-related mapping value. =2 indicates a temperature-related mapping value. =3 indicates a humidity-related mapping value. =4 indicates the dust concentration correlation mapping value), where... The calculation formula is: ; Indicates the first Environmental data in the first Influence factors of each image region; For the first The image region and the first Core pixel feature values ​​related to environmental data ( The time represents the average brightness of the region. The time represents the complexity of the region texture. Time represents the blurriness of the region image. (Time is the area pixel contrast). For all regions The average value of core pixel features related to environmental data; For the first Auxiliary pixel feature values ​​of each image region (such as...) Time represents the regional grayscale variance. (At the region edge density). For all regions The average value of the relevant auxiliary pixel features of the environmental data: The weighting coefficients for core pixel features ( hour , hour , hour , hour The core features, determined through extensive training with numerous samples, are used to ensure their dominant role in influencing the environment and comprehensively reflect the actual impact of environmental data on the image quality of the area. Further, the 4096×3072 image data of the hanging clamp is divided into (4096÷64)×(3072÷64)=64×48=3072 non-overlapping image regions. The region containing the bolt connection, due to its complex pixel features, is adaptively adjusted to a 32×32 pixel fine region, resulting in a total of 3072+10×3=3102 image regions (the fine region is calculated as equivalent to 3 standard regions). Then, Kriging interpolation is used to obtain the influence factor of each type of environmental data in each image region (e.g., near the top of the image). The area is directly exposed to strong sunlight, and the influence factor of the illumination data, after normalization, is 0.92; located at the metal bolt connection. In this region, due to the thermal conductivity characteristics of metal surfaces, the normalized influence factor of temperature data is 0.78; near the substation perimeter wall... In the region, dust deposition is relatively high, and the normalized influence factor of dust concentration is 0.85. Then, a 3102×4 environmental correlation mapping matrix M is formed, where M(100,1) = 0.92×( (Average brightness of region pixels / Average brightness of the overall image) = 0.92 × 1.2 = 1.104 (lighting-related mapping value); M(1205, 2) = 0.78 × ( Region texture complexity / overall image texture complexity) = 0.78 × 1.5 = 1.17 (temperature-related mapping value); M(3000, 4) = 0.85 × ( (Regional pixel contrast / overall image contrast) = 0.85 × 0.8 = 0.68 (dust-related mapping value).

[0015] In an optional embodiment, step S104 includes: obtaining a reference brightness value for each image region under standard illumination conditions. (Based on a large number of standard images of hardware without illumination interference, the average brightness of each type of image region is statistically analyzed and used as the baseline brightness value for that type of region); then, based on the illumination deviation calculation model, i.e. Calculate the illumination deviation value caused by uneven illumination in each image region, where, Indicates the first The reference brightness of each image region (in this embodiment, the standard region is set to 180 and the fine region is set to 200). This refers to the illumination deviation rate; For light intensity data, lux (standard light intensity); Indicates the first The illumination intensity correlation mapping value for each image region; A positive value indicates that the image area is too bright, and a negative value indicates that the image area is too dark; further, first determine... The baseline brightness for the standard area is 180 (grayscale value, 0-255), and the baseline brightness for the 32×32 fine area is 200 (because critical areas require higher brightness and clarity); the standard illumination intensity is known. lux, light intensity data lux, illumination deviation rate for area( Illumination deviation value (Too bright); for the shadow area at the bottom of the image area , (Low brightness); for fine-grained areas , , (Too bright), thus clarifying the illumination deviation value of the image area.

[0016] In an optional embodiment, step S105 specifically includes: firstly, determining the baseline sharpness of the image region without humidity interference by using an image blur detection algorithm (such as Laplacian gradient value calculation). (The higher the gradient value, the higher the sharpness), then calculate the current sharpness of the image region. Obtaining the blur caused by humidity Then, the humidity-related mapping value is combined. ( The larger the value, the stronger the ambiguity effect on the area. This is achieved by adjusting the humidity ambiguity reduction factor. , The correlation model is defined by the following formula: ;in, Humidity-induced blurring For the first The baseline sharpness of each image region (35 for standard region, 45 for fine region). Current sharpness of the image region: For the first The humidity-related mapping values ​​for each image region are considered; and considering that the image blur caused by humidity has a Gaussian distribution (the degree of blur varies with the distance of the region's pixels from the center of humidity influence), the parameter mapping relationship of the deconvolution algorithm is used to... and Converted to humidity-induced blur reduction amount , A larger value indicates a higher degree of blurring that needs to be eliminated, usually expressed as a sharpness enhancement gradient value. Further, the baseline sharpness for the standard region of the hanging clamp image is first calculated as 35 (Laplacian gradient value), and 45 for the finer region. For damp outdoor areas near substations... area Current resolution , For the dry indoor area area , , The amount of ambiguity reduction due to humidity. (The gradient value needs to be increased by 24.7 to eliminate blur). (Slightly reduce blur); fine-grained areas of , , ,but (Focus on eliminating blurriness at bolted connections).

[0017] In an optional embodiment, step S106 specifically includes: first calculating the standard temperature conditions ( The baseline contrast of the image region below (Calculated using the variance of the grayscale histogram; the larger the variance, the higher the contrast.) Then, the current contrast of the image region at the actual temperature is calculated. , obtained the Contrast deviation caused by temperature in each image region ( (This indicates a decrease in contrast); then combined with temperature-related mapping values. ( A larger value indicates a stronger effect of temperature on contrast reduction. Considering the physical property that temperature changes cause changes in the reflectivity of metal surfaces (thus affecting image contrast), a contrast compensation value is established. and The nonlinear mapping model is given by the following formula: Contrast compensation value The contrast improvement is expressed as a percentage, ensuring that the contrast of the compensated area does not exceed 1.2 times the baseline contrast (to avoid distortion caused by overcompensation). Further, a baseline contrast of 80 (grayscale variance) is first determined for the standard area, and 100 for the fine area; standard temperature... Temperature data is Temperature deviation For high-temperature areas middle Current contrast , For low-temperature shaded areas middle , , Then the contrast compensation value (Therefore, it needs to be improved) Contrast); (Slight compensation); Fine-grained area middle , , ,but (Contrast at the compensating bolt).

[0018] In an optional embodiment, step S107 includes: extracting edge sharpness data of the image region using an edge detection algorithm (Cany operator), including a baseline edge sharpness. and current edge sharpness (The ratio of edge pixel count to area; a higher ratio indicates higher sharpness), yielding the sharpness reduction rate caused by dust. Then, the dust concentration correlation mapping value was combined. ( The larger the value, the stronger the impact of dust on sharpness reduction. Considering the unevenness of dust coverage (such as particulate dust causing a sudden drop in sharpness in certain areas), an exponential mapping model is established, with the following formula: ;in, For the rate of decrease in sharpness; The baseline edge sharpness for the image region is 0.3 for the standard region and 0.4 for the fine region. The current edge sharpness of the image region; For the first Dust correlation mapping values ​​for each region. The value range is 0-1. The closer the value is to 1, the higher the required sharpness to be restored, and the more certain the restored image quality must be. (Avoid over-sharpening). Furthermore, determine the baseline edge sharpness for the standard area as... The fine-grained range is 0.4. For areas with high dust concentrations... area ,current , For those with less dust area , , The sharpness restoration factor is... (Needs to be restored) (Clarity) (Slight recovery); Fine-grained areas middle , , , (Restore the clarity of the bolt area).

[0019] This embodiment, through region division and influencing factor calculation, can accurately assess local environmental interference. By correlating mapping values, it explores the impact of various environmental factors on different regions of the image. By calculating illumination deviation values, humidity-induced blur reduction, contrast compensation values, and sharpness restoration coefficients, it precisely adjusts image data from multiple aspects such as brightness, sharpness, and contrast. It can specifically solve problems such as image noise, shadows, and distortion caused by illumination fluctuations, occlusion, and inconsistent shooting angles, thereby improving the accuracy of standardized image data and thus improving the accuracy of subsequent identification of corrosion defects in substation fittings.

[0020] In this embodiment, the image data is adjusted based on the illumination deviation value, contrast compensation value, humidity-induced blur reduction amount, and sharpness restoration coefficient of each image region to determine the standardized image data of the substation fittings, including steps S1081 to S1083. Step S1081: Based on the illumination deviation value of each image region, the brightness of each image region is corrected to determine the corrected brightness value for each image region. Step S1082: Based on the contrast compensation value, humidity-induced blur reduction amount, and sharpness restoration coefficient of each image region, the sharpness and contrast of each image region are jointly optimized to determine the comprehensive optimized quality value for each image region. Step S1083: Based on the corrected brightness value and comprehensive optimized quality value of each image region, the pixel coordinates of the image data are arranged to determine the standardized image data of the substation fittings.

[0021] In an optional embodiment, step S1081 specifically includes: obtaining the original average brightness value of each image region. (Obtained by the arithmetic mean of the brightness of all pixels within the statistical area), combined with the illumination deviation value calculated in step S103. First through The calibration benchmark is calculated after eliminating extreme values ​​to avoid overall calibration deviation caused by extreme values. The benchmark brightness is used as the target during calibration, with 180 for the standard area and 200 for the fine area, using a linear mapping formula: ;in, This is the effective average brightness value calculated after removing extreme values. The original brightness is adjusted to be close to the reference brightness range, while ensuring that the corrected brightness value falls within the grayscale range of 0-255 (any values ​​exceeding this range are truncated). After completing the brightness correction for a single area, the corrected average brightness value for that area is calculated. and verify (Grayscale value), if it does not meet the requirements, the mapping coefficients are readjusted (e.g., the correction amplitude is reduced) until the accuracy requirements are met, finally obtaining the corrected brightness value for each region (including the corrected brightness of each pixel within the region and the average corrected brightness of the region). Furthermore, for standard regions... Original average brightness (Grayscale value), Illumination deviation value Reference brightness First, we analyzed the pixel brightness distribution within the area and found 12 brightness levels. The extreme value (accounting for 0.3% of the total number of pixels in the area) is obtained through... After the criteria are removed, the effective average luminance is recalculated. The corrected brightness is calculated using a linear mapping formula: taking a certain pixel within the region as an example. For example, (Rounded to 111). Average brightness of the corrected area. , (Due to severe overexposure in the original image, a single correction could not fully meet the requirements). The mapping factor was adjusted to 0.8, and the result was recalculated. , The coefficient was adjusted again to 0.5, and finally... ,satisfy ,Sure Area correction brightness value (average 178, pixel brightness range) Fine-grained areas Original average brightness , Reference brightness Remove 5 extremely bright pixels (brightness). After that, the effective average brightness After correction using the mapping formula, the average brightness of the region , Determine the corrected brightness value (average 198, pixel brightness range) ).

[0022] In an optional embodiment, step S1082 specifically includes: first, based on the humidity-induced blurring reduction amount in step 105... And the sharpness restoration coefficient in step 106 An adaptive deconvolution-unsharpened mask joint algorithm is used to optimize the sharpness of image regions. The deconvolution stage is based on... Determine the deconvolution kernel size ( use nuclear, use nuclear, use (Nuclear), to eliminate Gaussian blur caused by humidity; and in the unsharpening mask stage according to Adjust mask intensity ( Use a weak mask Use a central mask. (Use a strong mask) to enhance edge sharpness. Then, in the contrast optimization stage, based on the contrast compensation value from step 106... The image region's grayscale range is divided into three segments: dark (0-80), mid-range (81-170), and bright (171-255). Based on... Adjust the stretching ratio of each segment (focus on stretching the central grayscale segment in areas with high contrast compensation requirements to highlight hardware details). After optimization, calculate the overall optimized quality value using a multi-dimensional quality assessment model. Its model formula is: ;in, , , For the indicator weights (and, , , (Clarity has the highest weight because corrosion recognition is most sensitive to edge details). The baseline sharpness for the image region (35 for standard region, 45 for fine region). The baseline contrast for the image region (80 for standard region, 100 for fine region). The quasi-edge percentage of the image region is used as the basis (0.3 for standard regions, 0.35 for fine regions). This comprehensive optimization quality value fuses sharpness (Laplacian gradient value). Contrast (grayscale variance) Edge integrity (the percentage of edge pixels detected by Canny edge detection) The three indicators are obtained by weighting after normalization, with values ​​ranging from 0 to 1. Values ​​closer to 1 indicate higher overall regional quality. Furthermore, the standard region... middle (use (Deconvolution kernel) (Medium-intensity unsharpened mask). After deconvolution, the Laplacian gradient value of the region increased from 22 to 42; after unsharpening the mask, the gradient value further increased to 46. Contrast optimization: Original grayscale variance By segmented stretching (stretching ratio of 1.3 for the middle grayscale range, 0.8 for the dark areas, and 1.0 for the bright areas), the grayscale variance was optimized. (satisfy Due to the low initial contrast ratio, the actual improvement was 68. The final overall optimized quality score is: Normalization ( )for (Truncation to 1) Normalization (benchmark) )for , (Edge pixel ratio, After normalization, it becomes Weighted average (High overall quality). Fine-grained area. middle (7×7 deconvolution kernel) (Medium intensity mask) After optimization (Base resolution 45, normalized to 1) (Base contrast ratio 100, normalized to 0.92). (benchmark) (Normalized value 0.875), overall quality value (Extremely high overall quality).

[0023] In an optional embodiment, step S1083 specifically includes: establishing an image region-pixel coordinate mapping table, and determining the range of pixel coordinates for each image region in the image data (e.g., ...). Corresponding coordinates (0-63, 0-63). (Corresponding to (320-351, 480-511)). During integration, the corrected brightness value from step S1081 is first fused with the optimized sharpness and contrast features from step S1082 at the pixel level. The final brightness of each pixel is adjusted by combining the corrected brightness value with the edge weights optimized for sharpness (edge ​​pixel brightness fine-tuning +5%, enhancing edge visibility), while ensuring that the pixel brightness difference at the splicing points of adjacent areas is ≤3 (grayscale value) to avoid splicing gaps. After integrating all areas, "global color equalization processing" is performed using the Adaptive Histogram Equalization (CLAHE) algorithm, limiting the contrast enhancement (clip limit=2.0) to eliminate color deviations between areas. Subsequently, the key indicators of the standardized image are verified: overall average brightness 175-185 (grayscale value), global contrast variance 75-85, and comprehensive quality value of key areas (such as bolt connections) ≥0.9. Once all indicators meet these standards, the image is determined to be standardized image data. Furthermore, through an image region-pixel coordinate mapping table, the... (Coordinates 0-63, 0-63, average corrected brightness 178) and adjacent (Coordinates 0-63, 64-127, average corrected brightness 182) stitched together, calculate the pixel brightness difference at the stitching point (coordinates 0-63, 63-64): Edge pixel brightness 180-185, Edge pixel brightness 183-188, maximum difference 3, meets the requirements; The brightness of edge pixels (such as the bolt threads) at the bolt connection (coordinates 320-351, 480-511) was slightly adjusted by +5%, increasing from 198 to 208, enhancing thread details. After processing with the CLAHE algorithm, the overall average brightness of the image is 181 (grayscale value, within the range of 175-185), and the global contrast variance is 82 (within the range of 75-85); at the bolt connection ( - The average comprehensive quality value of the area is 0.935 (≥0.9), with no overexposed / underexposed areas, no splicing faults, and a clear distinction between the silver-white of the galvanized layer and the yellowish-brown of the slight local rust. The final standardized image data of the suspension clamp is generated.

[0024] This embodiment corrects brightness by adjusting the illumination deviation value, and then optimizes sharpness and contrast by adjusting the contrast compensation value, humidity-induced blur reduction amount, and sharpness restoration coefficient. It comprehensively considers the impact of various environmental factors on image quality, which can more comprehensively improve the visual effect of the image, improve the accuracy of subsequent visual features, and thus improve the accuracy of subsequent identification of corrosion defects in substation fittings.

[0025] Step S2: Extract visual features of substation fittings from standardized image data, and determine the target type and installation location of the substation fittings based on the visual features. In this embodiment, extracting visual features of substation fittings from standardized image data and determining the target type and installation location of the substation fittings based on the visual features includes: inputting standardized image data into a pre-trained convolutional neural network to extract edge features, texture features, and color distribution features of the substation fittings to determine their visual features; and classifying the visual features of the substation fittings based on a preset support vector machine algorithm to determine the target type and installation location of the substation fittings.

[0026] In one alternative embodiment, a pre-trained feature extraction model based on a convolutional neural network (CNN) is used to process the standardized image data. The convolutional neural network comprises 12 convolutional layers, 4 pooling layers, and 2 fully connected layers. The convolutional layers use 3×3 convolutional kernels and extract edge features, texture features, and color distribution features of substation fittings through the ReLU activation function, i.e., extracting the visual features of the substation fittings. The pooling layers employ max pooling to reduce the dimensionality of the feature map while preserving key features, thus reducing computational cost. The fully connected layers fuse the extracted visual features into a global feature vector with a dimension of 1024, used to comprehensively represent the visual information of the fittings. Subsequently, the Support Vector Machine (SVM) algorithm is used to classify the extracted 1024-dimensional visual feature vector. For substation fitting type classification, the SVM training dataset contains visual feature samples of 100,000 sets for 15 common substation fittings, such as suspension clamps, tension clamps, and connecting fittings. For substation fitting installation location classification, the training dataset contains visual feature samples of fittings for 8 typical installation locations, such as the bus side, line side, and transformer outlet side, totaling 80,000 sets. The similarity between the input feature vector and the feature vectors of samples from each category is calculated using the Support Vector Machine (SVM) algorithm. The category with the highest similarity is determined as the target type and installation location of the substation fitting. Further, standardized image data of five sets of suspension clamp fittings are input into a CNN. The first three convolutional layers of the CNN extract edge features of the fittings, such as the curved edges of the clamps; the fourth to sixth convolutional layers extract texture features of the fitting surface, such as the galvanized texture; the seventh to ninth convolutional layers extract color distribution features, such as the silvery-white color of normal galvanized fittings and possible yellowish-brown rust marks; the tenth to twelfth convolutional layers further fuse these local features. After pooling, the fully connected layer outputs a 1024-dimensional global feature vector. This 1024-dimensional feature vector is then input into the SVM. The classification model (i.e., the Support Vector Machine (SVM) algorithm) shows that, in the classification of fitting types, the cosine similarity between this feature vector and the feature vector of the suspension clamp sample is 0.92, which is much higher than the similarity with other types of fittings such as tension clamps (0.65) and connecting fittings (0.58). Therefore, the target type of the substation fitting is determined to be a suspension clamp. In the classification of installation location, the cosine similarity between this feature vector and the feature vector of the installation location sample on the line side is 0.88, which is higher than the similarity with other locations such as the busbar side (0.72) and the transformer outlet side (0.69). Finally, the installation location of this suspension clamp is determined to be on the line side.

[0027] This embodiment extracts edge features, texture features, and color distribution features through a convolutional neural network, making the visual features more accurate and comprehensive. Then, the visual features are classified using a support vector machine algorithm, which can quickly and accurately determine the target type and installation location, providing an accurate data foundation for subsequent corrosion defect identification and improving the accuracy of corrosion defect identification of substation fittings.

[0028] Step S3: Based on the target type, installation location, and several environmental data of the substation fittings, perform node retrieval on the preset substation fitting knowledge graph to determine the environmental corrosion-related knowledge groups of the substation fittings.

[0029] In this embodiment, based on the target type, installation location, and several types of environmental data of the substation fittings, a node retrieval is performed on a preset substation fitting knowledge graph to determine the environmental corrosion-related knowledge group of the substation fittings, including steps S301 to S306. Step S301: Based on the target type and installation location of the substation fittings, the node layer of the preset substation fitting knowledge graph is retrieved to obtain the substation fitting target type node corresponding to the target type and the substation fitting installation location node corresponding to the installation location. Step S302: Based on the substation fitting target type node and substation fitting installation location node, the node association relationship of the relationship layer of the substation fitting knowledge graph is retrieved to determine the type-related node set corresponding to the target type and the location-related node set corresponding to the installation location, and the initial knowledge node group of the substation fittings is determined based on the type-related node set and the location-related node set. Step S303: Based on a preset matching degree function, the matching degree between each type of environmental data and each environmental node in the node layer of the substation fitting knowledge graph is calculated; the environmental nodes are filtered based on the matching degree to determine the first environmental node corresponding to each type of environmental data. Step S304: Based on the first environmental node, construct an environmental knowledge node group for substation hardware; and retrieve the relationship layer of the substation hardware knowledge graph to determine the corrosion association weight of each first environmental node. Step S305: Based on the corrosion association weight of each first environmental node, and combined with the preset environmental data influence coefficient for each type of environmental data, determine the node association strength value between any associated node in the initial knowledge node group and any first environmental node in the environmental knowledge node group. Step S306: Based on the node association strength value, filter the initial knowledge node group and the environmental knowledge node group to determine the environmental corrosion association knowledge group for substation hardware.

[0030] It should be noted that the pre-constructed substation fitting knowledge graph includes an entity layer, a relationship layer, and an attribute layer. The entity layer covers entities such as substation fitting types (e.g., suspension clamps, tension clamps), installation locations (e.g., line side, busbar side), environmental factors (e.g., light, humidity), corrosion types (e.g., pitting, uniform corrosion), and weak points (e.g., clamp bolt connections, mounting plate holes). The relationship layer defines the associations between entities, such as "suspension clamp - installed on - line side," "line side environment - prone to - pitting," and "suspension clamp - weak point - bolt connection." The attribute layer contains the specific attributes of each entity, such as the material attribute of the suspension clamp (galvanized steel), the humidity range attribute of the line side environment (50%-80%RH), and the characteristic attributes of pitting (localized pitting, yellowish-brown color). It adopts a combination of top-down and bottom-up approaches, first defining the top-level entity and relationship framework, then supplementing entity attributes and relationships by crawling power industry literature, standards and specifications, and historical inspection data, and finally constructing a substation hardware knowledge graph.

[0031] In an optional embodiment, step S301 specifically includes: using the target type (suspension clamp) and installation location (line side) of the substation fitting as search keywords, locating the corresponding nodes in the entity layer of the substation fitting knowledge graph, and recording them as the substation fitting target type node and the substation fitting installation location node.

[0032] In an optional embodiment, step S302 specifically includes: starting with the substation fitting target type node and the substation fitting installation location node, using a breadth-first search algorithm (BFS) to retrieve directly related first-level entity nodes (the association relationship is defined through a relation layer, such as "belongs to", "installed in", "has", etc.), and satisfying the following two conditions during the retrieval process: first, the association relationship between the node and the substation fitting target type node or the substation fitting installation location node is "directly related" (without intermediate nodes); second, the node attributes must be related to the substation fitting corrosion analysis (such as material, structure, weak parts, etc., excluding irrelevant attribute nodes such as "manufacturer" and "purchase date"); thereby forming a type-related node set corresponding to the target type and a location-related node set corresponding to the installation location, which serve as the initial knowledge node group for the substation fitting. Further, taking the fitting target type as the suspension clamp (substation fitting target type node... The installation location is on the line side (the node where the substation hardware is installed). For example, using BFS retrieval, to... Starting with the node, retrieve directly related nodes: associate "galvanized steel" (material node) through the "material attribute" relationship. () is associated with "bolted connections" (structural nodes) through the "structural composition" relationship. "Hanging plate" (structural node) (The "weak point" is used to link the "bolt connection" (weak point node) to the "weak point".) ), forming a set of type-related nodes ;by Starting with the node, retrieve directly related nodes: associate "outdoor exposure" (environmental feature node) through the "environmental characteristics" relationship. "Frequent wind and rain erosion" (environmental characteristic node) (220kV) is associated with the voltage level through the "voltage level adaptation" relationship. ), forming a set of location-related nodes Merge the two node sets, remove duplicate nodes (no duplicates), and obtain the initial knowledge node group. There are a total of 9 nodes, all of which are directly related to corrosion analysis.

[0033] In an optional embodiment, steps S303 and S304 specifically include: first, standardizing the environmental data (converting it to a normalized value in the range [0, 1]); then, calculating the matching degree between each type of environmental data and each environmental node in the node layer of the substation fitting knowledge graph using a matching degree function; including environmental nodes with a matching degree ≥ 0.8 in the environmental knowledge node candidate set, and finally obtaining the first environmental node corresponding to each type of environmental data; setting the matching degree function as: ;in, For standardized environmental data, ; For the first environmental nodes The midpoint of the corresponding numerical interval (e.g.) interval After standardization ); For the first environmental nodes Corresponding to the width of the numerical range (Ensure that at the boundaries of the numerical interval) (Conforms to fuzzy membership degree distribution) For the first environmental nodes The association weights with corrosion (obtained through pre-training, such as environmental nodes related to salt spray) Temperature-type environmental nodes Furthermore, regarding humidity... RH ( ),temperature ( Dust concentration ( Salt spray concentration ( Environmental data, humidity mapping is as follows: medium-high humidity nodes The numerical range is RH, If RH falls within the interval, the matching degree is calculated as follows: ,humidity RH ( Mapped to Temperature mapping is: mesothermal environment node The interval is , If it falls within the interval, the matching degree is 0.85, and it is mapped to... Dust concentration is mapped as: medium-high dust nodes The interval is , If it falls within the interval, the matching degree is 0.9, and it is mapped to... Salt spray concentration is mapped as: low salt spray node The interval is , If it falls within the interval, the matching degree is 0.88, and it is mapped to... After screening, environmental knowledge node groups are formed. These are all primary environmental nodes that directly affect corrosion; subsequently, the relationship layer of the substation hardware knowledge graph is searched to determine the corrosion association weight of each primary environmental node; among them, Medium galvanized steel ( )and Medium and low salt spray nodes Based on the corrosion sensitivity relationship, its corrosion correlation weight is 0.75; middle With medium to high humidity ( ): Through the oxidation acceleration relationship, its corrosion correlation weight is 0.8; bolted joints ( )and Based on the crevice corrosion relationship, the corrosion correlation weight is 0.9; With medium and high dust ( ): Corrosion is accelerated by dust accumulation, with a rust correlation weight of 0.85.

[0034] In an optional embodiment, steps S305 and S306 specifically include: setting the node association strength threshold to 0.6; Medium galvanized steel ( )and Medium and low salt spray Based on corrosion sensitivity, with a correlation weight of 0.75, combined with... If the environmental data influence coefficient is 0.8, then the correlation strength is calculated as follows: ; With medium to high humidity ( ): Through oxidation acceleration relationship, the correlation weight is 0.8. The environmental impact coefficient of RH is 0.7, indicating a strong correlation. (Below the threshold of 0.6, not retained for now); Bolted connections ( )and Based on the crevice corrosion relationship, the correlation weight is 0.9. Impact 0.7, Correlation Strength (≥0.6, retained); With medium and high dust ( The correlation between dust accumulation and accelerated corrosion is weighted at 0.85. Impact 0.8 (≥0.6, retained); the initial knowledge node group and the environmental knowledge node group are screened by the node association strength threshold to construct the environmental corrosion association knowledge group. , containing triples: >、< >、< > etc., a total of 8 effective triplets.

[0035] This embodiment determines the initial knowledge node group through relational layer retrieval, taking into account the structural and installation location factors of the substation fittings themselves. Then, the first environmental node is filtered through the matching degree function, so that the first environmental node can fully highlight the influence of environmental data. Then, the inherent relationship between fitting type, installation location and environmental factors is comprehensively considered through the node association strength value, so as to incorporate domain professional knowledge, thereby comprehensively capturing corrosion patterns and providing an accurate data foundation for the subsequent generation of targeted corrosion impact paths, thereby improving the accuracy of substation fitting corrosion defect identification.

[0036] Step S4: Based on the environmental corrosion association knowledge group, construct several corrosion influence paths for substation fittings, and extract several key corrosion knowledge nodes from the substation fitting knowledge graph for the corrosion influence paths. Then, based on the key corrosion knowledge nodes, construct the corrosion association knowledge of substation fittings. In this embodiment, based on the environmental corrosion association knowledge group, construct several corrosion influence paths for substation fittings, and extract several key corrosion knowledge nodes from the substation fitting knowledge graph for the corrosion influence paths. Then, based on the key corrosion knowledge nodes, construct the corrosion association knowledge of substation fittings, including steps S401 to S406. Step S401: Starting from the first environmental node and ending at the substation fitting target type node, perform path search on the substation fitting knowledge graph based on a preset path search algorithm and preset path constraints to construct several corrosion influence paths for substation fittings. Step S402: Based on the relationship layer of the substation fittings knowledge graph, determine the path weight of each corrosion impact path; based on the environmental data corresponding to the first environmental node of each corrosion impact path, determine the environmental data deviation rate of each corrosion impact path. Step S403: Based on the environmental data deviation rate and path weight of each corrosion impact path, combined with the environmental data influence coefficient and corrosion association weight, determine the corrosion association strength of each corrosion impact path. Step S404: Based on the number of corrosion impact paths and the corrosion association strength of each corrosion impact path, filter the corrosion impact paths to obtain several key corrosion impact paths. Step S405: Count the frequency of each knowledge node in the key corrosion impact paths, and based on a preset frequency threshold, filter several key corrosion knowledge nodes from the key corrosion impact paths. Step S406: Based on each key corrosion influence path and the corrosion correlation strength of the key corrosion influence path, and combined with key corrosion knowledge nodes, construct a corrosion correlation network; based on the preset knowledge graph reasoning rules, reason about the corrosion correlation network to construct corrosion correlation knowledge of substation fittings.

[0037] In an optional embodiment, steps S401 and S402 specifically include: starting from the first environmental node and ending at the target type node of the substation fittings, using a directed graph path search algorithm (i.e., a preset path search algorithm), setting path constraints including two conditions: first, all relationships in the path are "corrosion-related influence relationships" (such as "accelerate," "cause," "induce," etc.); second, the path length is ≤3 (source node → intermediate node → target node, to avoid logical breaks due to excessively long paths), searching the substation fittings knowledge graph for all directed paths that satisfy "environmental node → intermediate node → corrosion influence → T", obtaining several candidate paths. Simultaneously, through the relationship layer of the substation fittings knowledge graph, the node association strength values ​​in the candidate paths are multiplied to obtain the total path weight. Paths with a total path weight ≥0.4 are selected to obtain several corrosion influence paths, which are then arranged in descending order of weight to form a corrosion influence path set. Then, the environmental data corresponding to the first environmental node of each corrosion influence path, combined with the standard threshold corresponding to the environmental data, is used to calculate the environmental data deviation rate of each corrosion influence path. Further, using... node As the source node, For the target node, the search path is: Path 1: (Low salt spray) → (Galvanized steel) → (Suspension clamp), the relationship chain is "corrosion sensitivity → material composition", path weight = 0.6 ( - Association strength) × 0.9 - Association strength (weight of "material composition" relationship) = 0.54 (≥0.4, retained); Corresponding salt spray concentration Standard threshold Deviation rate Path 2: (Medium to high humidity) → (Bolt connection) → The relationship chain is "crevice corrosion → weak part belongs to", path weight = 0.63 × 0.85 = 0.535 (≥ 0.4, retained); node Corresponding humidity RH, standard threshold Deviation rate Path 3: (Medium to high dust levels) → → The relationship chain is "dust accumulation accelerates corrosion → weak points belong to", path weight = 0.68 × 0.85 = 0.578 (≥ 0.4, retained); node Corresponding dust concentration Standard threshold Deviation rate Path 4: (Medium temperature) → → The relationship chain is "temperature-accelerated oxidation → material composition", with a correlation strength of 0.5. - ) × 0.9 = 0.45 (≥0.4, retain); node Corresponding temperature Standard threshold Deviation rate After screening, a set of corrosion-affected paths is formed: P = [path 3 (0.578), path 1 (0.54), path 2 (0.535), path 4 (0.45)]. There are a total of 4 corrosion paths.

[0038] In an optional embodiment, steps S403, S404, and S405 specifically include: the formula for calculating the corrosion correlation strength is as follows: ; Path weights calculated for step S402: This is the environmental data deviation rate; if the actual value is less than the standard value, then... ; This is the environmental data impact coefficient, reflecting the differences in the impact of different environmental data on corrosion. The corrosion association weights for environmental nodes; The maximum value of the corrosion association weight is used for normalization. After calculating the corrosion association strength, the corrosion association strengths are sorted from high to low, and the strength values ​​of the top preset position (usually 60%-80% of the total number of paths; if the total number of paths is 4, then the preset position is 3) are taken as the key association strengths, and the corresponding paths are the key corrosion influence paths. Then, the frequency of each knowledge node in the key corrosion influence path is counted. The higher the frequency, the stronger the association between the node and corrosion, and the more likely it is to be a core node affecting corrosion. A frequency threshold is preset (usually 50% of the number of key corrosion influence paths; if there are 3 key paths, then the threshold is 2). Nodes with a frequency ≥ the frequency threshold are selected as key corrosion knowledge nodes. If there are nodes with the same frequency and all close to the threshold (such as multiple nodes with a frequency of 1), a second screening is performed based on the direct association between the node and corrosion (such as the "bolt connection" node being directly associated with crevice corrosion, with higher priority than the "hanging plate" node) to ensure that the key nodes can accurately reflect the core influencing factors of corrosion. Furthermore, the corrosion association strength of path 1... Corrosion correlation strength of path 2 Corrosion correlation strength of path 3 Corrosion correlation strength of path 4 Sort by intensity from highest to lowest: The preset position is 3, and the key correlation strengths are determined to be 0.925, 0.9, and 0.695. The key corrosion influence path is path 3. → → ), Path 1 ( → → ), Path 2 ( → → ); then count the frequency, nodes (Medium-high dust): Appears only in path 3, frequency = 1; node (Bolt connection): Appears in path 3 and path 2, frequency = 2; Node (Suspension clamp): Appears in all 3 paths, frequency = 3; Node (Low salt spray): Occurs only on path 1, frequency = 1; node (Galvanized steel): Appears only in path 1, frequency = 1 node (Medium-high humidity): Appears only in path 2, frequency = 1; frequency threshold = 3 × 50% = 1.5 (rounded to 2), filter nodes with a frequency ≥ 2: (Frequency 2) (Frequency 3); then a second screening: although , Node frequency 1, but The path with the highest strength is 3. Path 1, corresponding to the second strongest intensity, is considered in conjunction with the direct correlation between nodes and corrosion. Corrosion caused by dust accumulation (related to salt spray corrosion), supplementary screening , Key knowledge nodes for corrosion; the final set of key knowledge nodes for corrosion is as follows: In an optional embodiment, step S406 includes: constructing a "node-relationship-weight" corrosion association network with the key corrosion knowledge nodes as the core. Each node in the corrosion association network corresponds to a key knowledge node, and the directed edges between nodes correspond to the association relationships in the critical path (e.g., ...). → Corresponding to the relationship of "dust accumulation accelerates corrosion"), the weight of the edge is the corrosion correlation strength of the corresponding path (e.g. → The edge weight is 0.925. If multiple paths have the same node pair (skip if there is no same node pair), the average of the edge weights of these paths is calculated as the final weight. Isolated nodes (nodes without any edge connections, such as unrelated "voltage nodes") are deleted to ensure that each node in the corrosion association network is directly related to the corrosion impact. Knowledge graph reasoning rules are divided into three categories: 1. "Environment-Material-Corrosion Type" rules (e.g., "Medium-high dust + galvanized steel → localized corrosion"); 2. "Weak Point-Environment-Corrosion Risk" rules (e.g., "Bolt connection + medium-high humidity → increased crevice corrosion risk"); 3. "Corrosion Type-Development Rate" rules (e.g., "Pit corrosion + low salt spray → monthly corrosion depth 0.02mm"). A "forward chain reasoning algorithm" is used, with key corrosion nodes in the corrosion association network as trigger conditions, matching the corresponding rules in the rule base to infer potential association knowledge not directly presented in the corrosion association network (e.g., from "…"). (Medium to high dust levels) + (The matching rule for "(bolt connection)" leads to the inference that "bolt connections are prone to localized corrosion due to dust accumulation"). Finally, the potential related knowledge obtained through this inference is added to the initial knowledge node group, categorized and integrated according to "basic hardware information," "environmental impact information," "corrosion characteristic information," and "risk level information" to form the final corrosion-related knowledge. Further, the key corrosion knowledge nodes { Using} as the core, a corrosion-related network is constructed: nodes (Medium-high dust) and (Bolt connection): The edge corresponds to the "dust accumulation accelerates corrosion" relationship, with a weight of 0.925 (corrosion correlation strength of path 3); Node (Low salt spray) and (Galvanized steel, since M1 is an intermediate node of path 1, although it is not selected as a critical node, the connection must be retained): The edge corresponds to the "corrosion sensitivity" relationship, with a weight of 0.9 (the corrosion correlation strength of path 1); Node and (Suspension clamp): The edge corresponds to the "weak part belongs to" relationship, weight = (path 3 strength 0.925 + path 2 strength 0.695) / 2 ≈ 0.81; node and The edges correspond to the "material composition" relationship, with a weight of 0.9 (path 1 intensity); finally, delete unrelated nodes. , Classified as an "environmental impact node" Classified as "weak points / nodes" Classified as "Material Nodes" Classified as "Fitness Type Nodes"; the final network structure is clearly presented: Environmental Impact Nodes → Weak Part / Material Nodes → Fitting Type Nodes, with edge weights reflecting the strength of the association, providing clear topological relationships for subsequent reasoning; the key corrosion nodes in the corrosion association network trigger reasoning rules: Nodes (Low salt spray) + (Galvanized steel): Matching the "Environment-Material-Corrosion Type" rule "Low salt spray + galvanized steel → pitting corrosion", we can deduce the underlying knowledge that "the suspension clamp is made of galvanized steel, which is prone to pitting corrosion in low salt spray environments"; Node (Medium to high dust levels) + (Bolted Connection): Matching the rule "Weak Point - Environment - Corrosion Risk" "Bolted connection + medium to high dust → increased risk of localized corrosion", the underlying knowledge is deduced: "Bolted connections are prone to localized corrosion due to dust accumulation, requiring close monitoring"; Node (Bolt connection) + (Medium-high humidity): Matching the rule "Weak point - Environment - Corrosion risk" "Bolt connection + medium-high humidity → crevice corrosion risk", combined with a humidity of 65%RH (medium-high), the potential knowledge "low-level crevice corrosion risk exists at the bolt connection" is deduced; Supplement the potential knowledge to the initial knowledge node group, and integrate to form the final corrosion-related knowledge: Fitting basic information: Suspension clamp, material galvanized steel, installation location on the line side, weak point is the bolt connection; Environmental impact information: Medium-high dust, low salt spray, and medium-high humidity environments accelerate corrosion, while medium temperature environments have little impact on corrosion; Corrosion characteristic information: The main corrosion type is pitting corrosion (localized yellowish-brown pits), and there is a risk of localized corrosion and low-level crevice corrosion at the bolt connection; Risk level information: The corrosion risk level is medium, the pitting corrosion development rate is about 0.02mm per month, and localized corrosion requires quarterly inspection and monitoring.

[0039] This embodiment searches for corrosion impact paths from the environment to the corrosion of substation fittings. Then, it calculates the corrosion correlation strength using environmental data deviation rate, path weight, environmental data influence coefficient, and corrosion correlation weight. This allows for the selection of key corrosion impact paths that best reflect the corrosion patterns of substation fittings under the current environment. Next, it selects key corrosion knowledge nodes and constructs a corrosion correlation network. Finally, it uses knowledge graph reasoning rules to infer the corrosion correlation network, accurately obtaining the corrosion correlation knowledge of substation fittings. This improves the reliability of assessing the corrosion status of substation fittings and effectively solves the problems of missed detections and misjudgments caused by the lack of domain knowledge support in traditional single-vision methods, thus improving the accuracy of identifying corrosion defects in substation fittings.

[0040] Step S5: Based on visual features and corrosion correlation knowledge, perform path identification on the preset power defect identification map model, determine the effective path group of substation hardware, and combine visual features and corrosion correlation knowledge to confirm the defects of substation hardware, and obtain the corrosion defect identification result.

[0041] In this embodiment, based on visual features and corrosion correlation knowledge, path identification is performed on the preset power defect identification graph model to determine the effective path group of substation hardware. Then, combining visual features and corrosion correlation knowledge, defects in the substation hardware are confirmed to obtain corrosion defect identification results, including steps S501 to S505. Step S501: Based on the target type and installation location of the substation hardware, label matching is performed on the graph model nodes of the preset power defect identification graph model to filter several graph model nodes, forming graph model node subgroups. Step S502: A matching degree matrix is ​​constructed between visual features and graph model node subgroups. Based on the matching degree matrix and the power defect identification graph model, the graph model nodes in the graph model node subgroups are filtered to obtain candidate graph model node groups. Step S503: Based on corrosion association knowledge and node subgroups, analyze the attribute information of the graph model edges in the power defect identification graph model, select several graph model edges to form graph model edge subgroups; and perform association strengthening processing on each graph model edge in the graph model edge subgroups based on corrosion association knowledge to obtain the graph model edge attribute matrix. Step S504: Based on the candidate graph model node groups and graph model edge attribute matrix, perform path identification on the power defect identification graph model to determine the effective path groups for substation hardware. Step S505: Based on the effective path groups of substation hardware, visual features, and corrosion association knowledge, confirm the defects of the substation hardware to obtain the corrosion defect identification results.

[0042] It should be noted that the pre-built power defect identification graph model is based on a large amount of historical data and expert experience, and is established through a complex knowledge graph foundation and graph neural network driven fusion architecture technology. It takes the corrosion defect scenario of substation fittings as its core, constructs a structured foundation through knowledge graph technology, and uses multi-layer graph neural networks (such as CompGCN and GAT) as the model backbone. It constructs a core training dataset through image data, structured data, and environmental data. The knowledge graph ensures the interpretability of the model (e.g., the defect association path is clearly traceable), while the graph neural network improves the identification accuracy for complex scenarios.

[0043] This embodiment filters graph model node subgroups and constructs a matching degree matrix to filter candidate graph model node groups, providing a suitable node foundation for path recognition. It then filters graph model edge subgroups and performs association enhancement processing to obtain a graph model edge attribute matrix, considering the influence of corrosion association knowledge on edges. Subsequently, using the candidate graph model node groups and the graph model edge attribute matrix, it filters out the effective path groups that best reflect the association patterns of corrosion defect characteristics of substation fittings under the current operating scenario. Finally, it uses the effective path groups, visual features, and corrosion association knowledge of substation fittings to confirm defects, accurately obtaining the defect recognition results and achieving corrosion defect recognition of substation fittings, thus improving the accuracy of corrosion defect recognition.

[0044] In this embodiment, a matching degree matrix between visual features and graph model node subgroups is constructed. Based on the matching degree matrix and the power defect identification graph model, the graph model nodes in the graph model node subgroups are screened to obtain candidate graph model node groups, including steps S5021 to S5027. Step S5021: Decompose the visual features to construct several feature sub-vectors. Step S5022: Calculate the cosine similarity between each feature sub-vector and the feature attribute vector of the graph model node; calculate the ratio between each feature sub-vector and the feature attribute vector of the graph model node to determine the attribute matching degree between each feature sub-vector and the graph model node. Step S5023: Based on the cosine similarity and attribute matching degree, determine the comprehensive matching degree between each feature sub-vector and the graph model node, and then construct a matching degree matrix between visual features and graph model node subgroups based on the comprehensive matching degree. Step S5024: Obtain the number of edges in the graph model of the power defect identification graph model, and determine the degree centrality of each graph model node in the graph model node subgroup by combining the number of incoming and outgoing edges of each graph model node in the graph model node subgroup. Step S5025: Determine the maximum matching degree of each graph model node in the graph model node subgroup based on the matching degree matrix. Step S5026: Perform a weighted summation of the degree centrality and maximum matching degree of each graph model node in the graph model node subgroup to determine the comprehensive index of each graph model node in the graph model node subgroup. Step S5027: Filter the graph model nodes in the graph model node subgroup based on the comprehensive index to obtain the candidate graph model node groups.

[0045] This embodiment constructs feature sub-vectors by decomposing visual features, and uses cosine similarity and attribute matching degree to build a matching degree matrix between visual features and graph model node subgroups, which can accurately measure the matching degree between visual features and graph model nodes. Then, by weighted summation using degree centrality and maximum matching degree, it comprehensively considers the matching degree of visual features and the topological structure information of graph model nodes, which can more accurately filter out graph model nodes related to visual features, providing an accurate data foundation for the subsequent acquisition of effective path groups, thereby improving the accuracy of identifying corrosion defects in substation fittings.

[0046] In this embodiment, defects in substation fittings are confirmed based on the effective path group, visual features, and corrosion association knowledge, resulting in corrosion defect identification. This includes steps S5051 to S5058. Step S5051: Based on the path defect type mapping relationship library in the power defect identification graph model, each effective path in the effective path group of the substation fitting is judged to determine the defect type of each effective path. Step S5052: Based on the defect type of each effective path, the frequency of occurrence of each defect type is determined, and the defect types are filtered based on the frequency of occurrence to determine a candidate defect type group. Step S5053: Based on corrosion association knowledge, the degree of fit between each defect type in the candidate defect type group and the corrosion association knowledge is calculated, and the corrected confidence level of each defect type is determined based on the degree of fit. Step S5054: Based on the corrected confidence level, the defect types are filtered to determine a confidence defect type group. Step S5055: Based on the power defect identification graph model, obtain the visual feature group corresponding to each confidence defect type in the confidence defect type group. Step S5056: Calculate the consistency score between the visual feature group and the visual features corresponding to each confidence defect type in the confidence defect type group, and filter the confidence defect type group based on the consistency score to obtain the defect type group to be confirmed. Step S5057: Based on the power defect identification graph model, obtain the path strength of each defect type to be confirmed in the defect type group, and calculate the confidence assessment value of each defect type to be confirmed by combining the corrected confidence of each defect type. Step S5058: Take the defect type with the highest confidence assessment value as the corrosion defect type of the substation fittings, and obtain the corrosion defect identification result by combining visual features, corrosion association knowledge, and effective paths.

[0047] This embodiment determines the defect types of effective paths, filters candidate defect type groups, and initially identifies possible defect types. Then, it calculates the fit between candidate defect types and corrosion association knowledge to determine the corrected confidence level, thereby filtering confidence level defect type groups and considering the correction of defect types by corrosion association knowledge. Subsequently, it further filters to obtain the defect type group to be confirmed through consistency scores, and determines the corrosion defect type of substation hardware with confidence assessment value, constructing corrosion defect identification results. It comprehensively utilizes path information, visual features, and corrosion association knowledge, improving the accuracy of substation hardware corrosion defect identification.

[0048] In an optional embodiment, step S501 specifically includes: constructing a tag matching rule based on the target type (suspension clamp) and installation location (line side) of the substation fittings. The rule is that the node attribute contains the tag "fitting type = suspension clamp" or "installation location = line side," and the node type belongs to "visual feature nodes" (such as edge feature nodes, color feature nodes, texture feature nodes) or "corrosion knowledge nodes" (such as corrosion type nodes, weak point nodes). Using the tag matching rule, several graph model nodes are extracted from the node layer of the power defect identification graph model to form graph model node subgroups. Further, in the power defect identification graph model, nodes containing the tag "fitting type = suspension clamp" are selected: edge feature nodes. (Attributes: Edge gradient threshold 15-30), Color feature nodes (Attributes: Normal zinc plating color RGB[240, 245, 250], rust color RGB[150, 100, 50]), texture feature nodes (Attributes: Galvanized texture density 10-15 lines / mm); Nodes with the label "Installation Location = Line Side": Weak point nodes. (Attributes: Coordinate range of bolted connections), Rust type nodes (Attribute: Pitting feature description), forming a graph model node subgroup .

[0049] In an optional embodiment, steps S5021 to S5027 specifically include: decomposing the visual features to construct several feature sub-vectors, including edge feature sub-vectors, color feature sub-vectors, and texture feature sub-vectors; then mapping them to the feature attributes of corresponding types of nodes in the graph model node subgroup (such as the "edge gradient threshold" attribute of edge feature nodes and the "RGB color range" attribute of color feature nodes), and using a combination of cosine similarity and attribute matching degree to determine the comprehensive matching degree between each feature sub-vector and the graph model node, thereby constructing a matching degree matrix between the visual features and the graph model node subgroup; the formula for calculating the comprehensive matching degree is: ;in, For feature vectors graph model nodes in a subgroup of graph model nodes Feature attribute vector Cosine similarity (values ​​0-1); For feature vectors Dimensions (such as edge feature vectors) ); For graph model nodes The effective range of feature attribute vectors (e.g.) Edge gradient threshold 15-30); For indicator functions, The value is 1 if the condition is met, otherwise it is 0. The attribute matching degree (value 0-1) reflects the proportion of feature sub-vectors falling within the effective range of feature attribute vectors; then, the number of graph edges of the power defect identification graph model is obtained. And combine the sum of the number of incoming and outgoing edges of each graph model node in the graph model node subgroup. Determine the degree centrality of each graph model node in the graph model node subgroup. Then, the maximum matching degree of each graph model node in the graph model node subgroup is queried through the matching degree matrix. Then, the composite index is calculated. The formula for the composite index is: ;in, and All are weighting coefficients. , ; For graph model nodes The number of incoming edges plus the number of outgoing edges; The number of edges in the graph model for power defect identification; For graph model nodes With all feature vectors The maximum matching degree is then determined; subsequently, the graph model nodes in the graph model node subgroup are filtered using a comprehensive index to obtain the candidate graph model node group. Furthermore, the edge feature subvectors are compared with... Edge gradient threshold attribute matching: Cosine similarity is 0.85, attribute matching degree (the proportion of dimensions with gradient values ​​in the sub-vector range of 15-30) is 0.92, and the overall matching degree = 0.85 × 0.92 = 0.782; color feature sub-vector and of Color range attribute matching: Normal galvanized color matching similarity 0.7, rust color matching similarity 0.88, attribute matching degree (rust color dimension proportion) 0.3, overall matching degree = 0.88 × 0.3 = 0.264; texture feature sub-vector and The matching of galvanized texture density attributes: similarity 0.82, attribute matching degree 0.88, overall matching degree = 0.82 × 0.88 = 0.7216; color feature sub-vector and The pitting feature description (color) attribute matching: similarity 0.9, attribute matching degree 0.3, overall matching degree = 0.9 × 0.3 = 0.27; finally, a matching degree matrix is ​​constructed. (Rows: edges, colors, textures; Columns:) ): Degree centrality calculation: Number of edges in the graph model 100, graph model node subgroup The sum of the number of incoming edges and the number of outgoing edges of each node in the graph model is: (8 items) (12 items) (7 items) (10 items) (15 items), with degree centrality values ​​of 0.08, 0.12, 0.07, 0.1, and 0.15 respectively. Therefore, the overall importance index (weight: degree centrality 0.4, matching degree 0.6) is calculated as follows: : 0.08×0.4 + 0.782×0.6 = 0.032 + 0.4692 = 0.5012; : 0.12×0.4 + 0.264×0.6 =0.048 + 0.1584 = 0.2064; : 0.07×0.4 + 0.7216×0.6 = 0.028 + 0.43296 = 0.46096; : 0.1 × 0.4 + 0 (no match) × 0.6 = 0.04; : 0.15×0.4 + 0.27×0.6 =0.06 + 0.162 = 0.222; Filtering yields: Average comprehensive index = (0.5012+0.2064+0.46096+0.04+0.222) / 5≈0.286, Threshold = 0.286×1.2≈0.343; Nodes with indices exceeding the threshold: After sorting, the top 80% are selected (both are included), and supplementary data is added. With connection (Although the exponent 0.222 is less than the threshold, it is a key node for defect judgment), thus obtaining the candidate graph model node group. .

[0050] In an optional embodiment, step S503 specifically includes: traversing the graph model edge set of the power defect identification graph model and filtering graph model edges related to corrosion association knowledge; the filtering criteria are: the "association description" attribute of the graph model edge contains keywords in the corrosion association knowledge (such as "pitting corrosion", "bolt connection", "galvanized steel"), and the two nodes connected by the edge both belong to the graph model node subgroup of step S501 (or one belongs to the node subgroup and the other belongs to the "defect judgment node"). For example, connecting "color feature nodes" - Rust type nodes Furthermore, edges whose relationship description is "color feature corresponds to pitting" meet the screening criteria and form a graph model edge subgroup. Next, for each graph model edge in the graph model edge subgroup, information related to the edge relationship in the corrosion association knowledge is extracted (such as "pitting corrosion development rate 0.02mm per month" and "medium risk of corrosion at bolt connections"). The knowledge association degree is calculated, and it is set as the semantic similarity between the edge relationship and the corrosion association knowledge (calculated through a pre-trained BERT model). The knowledge association degree is then combined with the original weights of the graph model edges (predefined in the power defect identification graph model) using the weight strengthening formula: ;in, For graph model edges The original weights; For graph model edges The text describing the relationship; For knowledge text related to corrosion; for and semantic similarity; The edge weights of the enhanced graph model are then calculated. Next, the edge attributes are updated to obtain the enhanced edge attribute matrix, where the matrix elements represent the enhanced graph model edge weights, ensuring a significant increase in the weights of edges highly correlated with corrosion knowledge. Further, edge subgroup selection is performed: edge subgroups are selected from the graph model edge set. ( The relationship describes the color feature corresponding to pitting, containing the keyword "pitting," and the connecting nodes are all in... Selected; ( The relationship description of texture anomalies corresponding to pitting corrosion, containing the keyword "pitting corrosion", is selected; ( The relationship description indicates that weak areas are prone to pitting corrosion, and the keywords "pitting corrosion weak areas" are included. ( The relationship describes the pitting type and its corresponding defect level. The node used for defect assessment is selected. Ultimately, this forms an edge subgroup. The original weights were 0.6, 0.55, 0.7, and 0.8, respectively. Then, edge attribute association strengthening was performed: The corrosion-related knowledge point is characterized by localized, dotted, yellowish-brown depressions. The semantic similarity to the edge relationship is 0.85 (knowledge relevance). The edge weights in the enhanced graph model are... ; Knowledge: Abnormal texture in galvanized steel may be accompanied by pitting corrosion; semantic similarity 0.78; enhanced graph model edge weights = ; Pitting corrosion is prone to occur at bolted connections where knowledge is weak; semantic similarity is 0.92; edge weights of the enhanced graph model = ; The knowledge point has a medium risk level of corrosion, corresponding to slight corrosion, with a semantic similarity of 0.88. The edge weights of the enhanced graph model are... Finally, construct the edge attribute matrix of the graph model. (OK Column: Edge Index (Diagonal lines represent the weights after enhancement, off-diagonal lines are 0). .

[0051] In an optional embodiment, step S504 specifically includes: using the candidate graph model node group and the graph model edge attribute matrix, employing a depth-first search (DFS) algorithm to perform path identification on the power defect identification graph model, with the rule that the path starting point is a "visual feature node" ( , The endpoint is the "defect judgment node" ( The path length is ≤3 (visual feature node → corrosion knowledge node → defect judgment node), and all edges in the path come from the graph model edge subgroup. Next, for each mined path, the path strength is calculated, which is the product of the weights of all strengthened graph model edges. A path strength threshold is set (70% of the maximum strength of all paths), and paths with strengths higher than the threshold are selected as valid paths. Furthermore, based on... and Three paths were discovered: : (Edge features) No direct edges (Invalid, no matching edges); : (Texture features) (Texture anomalies correspond to pitting, weight 0.979) (Type of corrosion) (Pitting corresponds to a defect level with a weight of 1.504) (Valid, length 3); : (Color characteristics, although not selected into the candidate node group, but...) starting point) (Color corresponds to pitting, weight 1.11) (Valid, supplementing color feature path); : (Weak points, supplementary nodes) (Weak areas are prone to pitting corrosion, weight 1.344) (Valid, supplementing weak points in the path); Path strength calculation is as follows: ; ; The filtering yielded: maximum intensity 2.021, threshold... If the strength of all three paths is higher than the threshold, then the path group is considered valid. .

[0052] In an optional embodiment, steps S5051 to S5058 specifically include: based on a predefined path defect type mapping relationship library, the mapping rules use "path node combination + edge relationship description" as the key and "defect type" as the value. For example, the path "texture feature node → rust type node (pitting corrosion) → defect judgment node" corresponds to "pitting corrosion", and the path "color feature node (yellowish brown) → rust type node (uniform rust) → defect judgment node" corresponds to "uniform rust", thus obtaining the defect type of the effective path; traversing each effective path in the effective path group, extracting the "node combination sequence" of the effective path (e.g., The node combination is ) and "set of edge relation descriptions" (such as "Texture anomalies correspond to pitting" The path defect type mapping database is used to perform precise matching on the "pitting corrosion corresponding to the defect level". If a completely matching mapping rule exists, the defect type of the valid path is directly determined; if only a partial match exists (such as consistent node combinations or similar edge relationship descriptions), the similarity between the path and the mapping rule is calculated using a "semantic similarity matching algorithm" (based on a pre-trained BERT model), and the defect type corresponding to the mapping rule with a similarity ≥ 0.85 is taken as the defect type; then, the frequency of occurrence of the defect type is counted, and a defect type frequency distribution table is constructed with "defect type name" as the row index and "occurrence frequency" and "corresponding path list" as columns. For example, "pitting corrosion" appears twice in 3 paths, and the corresponding path is... , If so, the record in that row of the table will be "pitting corrosion, 2 times". Then, a frequency threshold is set (usually 50% of the total number of valid paths; if the number of valid paths is 3, then the threshold = 1.5, rounded down to 2), and defect types with a frequency ≥ the threshold are filtered. If there are types with the same frequency and both below the threshold (e.g., two defect types each appear once), then the type with the higher total path intensity is added to the candidate defect type group, based on the sum of the path intensity corresponding to the defect type (the higher the path intensity, the higher the type credibility). Then, the rust-related knowledge (e.g., "the main rust type is pitting corrosion, manifested as localized pointy yellowish-brown depressions" and "there is a risk of localized rust at bolt connections") is broken down into text fragments in three dimensions: "rust type description," "rust characteristic description," and "risk location description." For each defect type in the candidate defect type group, the semantic similarity between the type name, feature description, and knowledge fragments in the three dimensions is calculated. Then, a weighted fusion is performed (weights: type description 0.4, feature description 0.3, location description 0.3) to obtain the fit. Next, an initial confidence level is set for each defect type (initial confidence level = sum of path strengths corresponding to the defect type / sum of path strengths of all defect types). The fit is then multiplied by the initial confidence level to obtain the corrected confidence level. A confidence threshold is set (usually 0.6), and types with a corrected confidence level ≥ the threshold are selected to form a confidence defect type group. For each confidence defect type in the confidence defect type group, the relevant information is extracted from the power defect identification graph model. The visual feature group corresponding to each type includes the feature parameter ranges of edge features (such as the "irregular edge gradient distribution" of pitting corrosion), color features (such as the "yellowish-brown RGB range"), and texture features (such as the "zinc-plated texture fracture area"). Then, the visual features (edge ​​feature sub-vectors, color feature sub-vectors, and texture feature sub-vectors) are matched with the visual feature group for consistency. A weighted sum is then performed (weights set to 0.3 for edges, 0.4 for colors, and 0.3 for textures) to calculate a consistency score. A consistency score threshold is set (usually 0.4-0.7), and types with scores ≥ the threshold are selected to form a group of defect types to be confirmed. Then, for each defect type to be confirmed in the confirmed defect type group, the following steps are performed: The confidence assessment value is calculated by weighted summation of three indicators: path strength contribution value, corrected confidence score, and consistency score. The path strength contribution value is calculated as: (sum of path strengths corresponding to the defect type to be identified) / (sum of total effective path strengths). The weights are: path strength contribution value 0.4, corrected confidence score 0.3, and consistency score 0.3. The confidence assessment values ​​are then sorted from highest to lowest, and the type with the highest value is selected as the core defect type. This is then combined with the corresponding corrosion association knowledge (e.g., corrosion location, development rate), visual features (e.g., proportion of matched feature dimensions), and effective paths (e.g., key support paths) to form the corrosion defect identification result. Furthermore, the effective path group... , ( (Edge relationships: texture anomalies correspond to pitting corrosion, pitting corrosion corresponds to defect levels): In the mapping library, a complete match rule is used: "texture feature node → pitting corrosion type node → defect judgment node → pitting corrosion type corrosion" to initially determine the defect type. "Pit corrosion"; ( Edge relationships: color corresponds to pitting corrosion, pitting corrosion corresponds to defect level): perfect matching rule "color feature node → pitting corrosion type node → defect judgment node → pitting corrosion type corrosion", initial judgment of defect type. "Pit corrosion"; ( Edge relationships: weak points are prone to pitting corrosion, pitting corrosion corresponds to defect level): perfect match rule "weak point node" Pitting type nodes Defect judgment node Pitting corrosion (localized), preliminary assessment of defect type "Local pitting corrosion"; ultimately, the defect type for each effective path is obtained: , , ; then, the frequency of occurrence was counted, "pitting corrosion": appeared 2 times, corresponding to the path (Intensity 1.472) (Intensity 1.670), total path intensity = 1.472 + 1.670 = 3.142; "Local pitting corrosion": occurs once, corresponding to the path (Intensity 2.021), total path intensity = 2.021; after screening, the frequency threshold = 2, "pitting corrosion" frequency = 2 ≥ threshold, directly selected; "local pitting corrosion" frequency = 1 < threshold, but the total path intensity 2.021 is relatively high (close to "pitting corrosion" 3.142), and belongs to the same "pitting" category as "pitting corrosion", so it is added to the selection; final candidate defect type group. { The relevant knowledge fragments for corrosion are as follows: Type description: "The main type of corrosion is pitting corrosion"; Feature description: "Manifests as localized yellowish-brown pits"; Location description: "There is a risk of localized corrosion at bolted connections"; The calculated fit is: "For pitting corrosion": similarity with type description = 0.95, similarity with feature description = 0.8, similarity with location description = 0.6; therefore, the fit = 0.95 × 0.4 + 0.8 × 0.3 + 0.6 × 0.3 = 0.38 + 0.24 + 0.18 = 0.8; "For localized pitting corrosion": similarity with type description = 0.95, similarity with feature description = 0.9 (including "localized," matching "localized pitting" in the knowledge), similarity with location description = 0.9 (matching "localized corrosion at bolted connections"). Overall fit = 0.95×0.4 + 0.9×0.3 + 0.9×0.3 = 0.38 + 0.27 + 0.27 = 0.92; Confidence correction: Sum of path strengths for all defect types = 3.142 + 2.021 = 5.163; Initial confidence for "pitting corrosion" = 3.142 / 5.163 ≈ 0.608, corrected confidence = 0.608×0.8 ≈ 0.486; Initial confidence for "localized pitting corrosion" = 2.021 / 5.163 ≈ 0.391, corrected confidence = 0.391 × 0.92 ≈ 0.359; Screening: Confidence threshold = 0.6. Although the confidence scores of both types are below the threshold after adjustment, "pitting corrosion" is closer to the threshold and is a high-frequency type. Considering the knowledge that "the main type of corrosion is pitting corrosion," the threshold is adjusted to 0.45, and both types are selected, forming the confidence defect type group. .

[0053] Construct visual feature groups: "Pit corrosion type" with edge gradient of 15-30, color RGB [150-170, 100-120, 50-70], and texture fracture density ≥2 locations / mm²; "Local pit corrosion type" with edge gradient of 18-25, color RGB [140-160, 90-110, 40-60], and texture fracture density ≥3 locations / mm². Calculate the consistency score: Edge feature sub-vector (256 dimensions): 180 dimensions fall within the range of 15-30, and 150 dimensions fall within the range of 18-25; Edge matching ratio for "pitting corrosion" = 180 / 256 ≈ 0.703, Edge matching ratio for "local pitting corrosion" = 150 / 256 ≈ 0.586; Color feature sub-vector (256 dimensions): 80 dimensions fall within [150-170, 100-120, 50-70], and 95 dimensions fall within [140-160, 90-110, 40-60]; Color matching ratio for "pitting corrosion" = 80 / 256 ≈ 0.3 12. Color matching ratio for "Local Pitting Corrosion" = 95 / 256 ≈ 0.371; Texture feature sub-vector (512 dimensions): 200 dimensions correspond to a texture fracture density ≥ 2 locations / mm², 180 dimensions correspond to ≥ 3 locations / mm²; Texture matching ratio for "Pitting Corrosion" = 200 / 512 ≈ 0.391, Texture matching ratio for "Local Pitting Corrosion" = 180 / 512 ≈ 0.352; Consistency score: "Pitting Corrosion" = 0.453; "Local Pitting Corrosion" = 0.430; Screening: Consistency score threshold = 0.4, both types are selected, pending confirmation of defect type group. .

[0054] The total strength of the effective path = 1.472 + 1.670 + 2.021 = 5.163; the calculated confidence assessment value is: "Pit corrosion": path strength contribution value = 3.142 / 5.163 ≈ 0.608; corrected confidence level = 0.486, consistency score = 0.453; confidence assessment value = 0.525; "Local pitting corrosion": path strength contribution value = 2.021 / 5.163 ≈ 0.391; corrected confidence level = 0.359, consistency score = 0.430; confidence assessment value = 0.393; therefore, when sorting and integrating the results, the assessment value of "pitting corrosion" 0.525 > 0.393, and it is determined to be the core defect type; combined with the corrosion association knowledge "local corrosion risk at bolt connections", visual characteristics "color matching yellowish-brown proportion 37.1%", and effective path " (Texture) (Color)”, final defect identification result: the suspension clamp has pitting corrosion defects, mainly concentrated at the bolt connection. The corrosion characteristics are local spot-like yellowish-brown depressions. The color and texture characteristics match well in the visual features. The defect development rate is about 0.02 mm per month.

[0055] Please refer to Figure 2 , Figure 2 This invention provides a schematic diagram of a substation fitting corrosion defect identification system, comprising: an image standardization module 6, a visual feature extraction module 7, a knowledge graph retrieval module 8, a corrosion association knowledge construction module 9, and a corrosion defect identification result acquisition module 10. The image standardization module 6 acquires image data of the substation fitting and several types of environmental data, and adjusts the image data based on the environmental data to determine standardized image data of the substation fitting. The visual feature extraction module 7 extracts visual features of the substation fitting from the standardized image data and determines the target type and installation location of the substation fitting based on the visual features. The knowledge graph retrieval module 8 determines the target type, installation location, and other relevant information of the substation fitting based on the visual features. The system uses a variety of environmental data to perform node retrieval on a pre-defined substation hardware knowledge graph, identifying environmental corrosion-related knowledge groups for the substation hardware. The corrosion-related knowledge construction module 9 is used to construct several corrosion impact paths for the substation hardware based on these environmental corrosion-related knowledge groups, extract several key corrosion knowledge nodes from the substation hardware knowledge graph based on these paths, and then construct corrosion-related knowledge for the substation hardware based on these key corrosion knowledge nodes. The corrosion defect identification result acquisition module 10 is used to perform path identification on a pre-defined power defect identification graph model based on visual features and corrosion-related knowledge, identify effective path groups for the substation hardware, and confirm defects in the substation hardware by combining visual features and corrosion-related knowledge, obtaining corrosion defect identification results.

[0056] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for identifying corrosion defects of a power transformation fitting, characterized in that, include: Image data of substation fittings and several types of environmental data are acquired, and the image data is adjusted based on the environmental data to determine the standardized image data of the substation fittings. Visual features of the substation fittings are extracted from the standardized image data, and the target type and installation location of the substation fittings are determined based on the visual features. Based on the target type, installation location, and several environmental data of the substation fittings, a node retrieval is performed on a preset substation fitting knowledge graph to determine the environmental corrosion-related knowledge groups of the substation fittings. Based on the environmental corrosion-related knowledge group, several corrosion influence paths of the substation fittings are constructed, and several key corrosion knowledge nodes of the substation fitting knowledge graph are extracted from the corrosion influence paths. Then, based on the key corrosion knowledge nodes, the corrosion-related knowledge of the substation fittings is constructed. Based on the visual features and the corrosion association knowledge, path identification is performed on the preset power defect identification map model to determine the effective path group of the substation hardware. The defects of the substation hardware are confirmed by combining the visual features and corrosion association knowledge to obtain the corrosion defect identification result.

2. The method for identifying corrosion defects of a power transformation fitting according to claim 1, characterized in that, The process of acquiring image data of substation fittings and several types of environmental data, and adjusting the image data based on the environmental data to determine standardized image data of the substation fittings includes: Image data of substation fittings and several types of environmental data are acquired, including: light intensity data, temperature data, humidity data, and dust concentration data; The image data is divided into several image regions, and the influence factor of each type of environmental data in each image region is calculated. Based on the environmental data and the influencing factors, calculate the light intensity data and the light intensity correlation mapping value of each image region, the temperature data and the temperature correlation mapping value of each image region, the humidity data and the humidity correlation mapping value of each image region, and the dust concentration data and the dust concentration correlation mapping value of each image region; Based on the light intensity correlation mapping value, and combined with the preset reference brightness of each image region, the light deviation value of each image region is calculated; Obtain the blur level of each image region and, in conjunction with the humidity-related mapping value, determine the amount of humidity-induced blur reduction for each image region; Obtain the contrast deviation value of each image region, and combine it with the temperature-related mapping value to determine the contrast compensation value of each image region; Extract edge sharpness data for each image region, determine the sharpness reduction rate for each image region based on the edge sharpness data, and combine the dust concentration correlation mapping value to determine the sharpness recovery coefficient for each image region; Based on the illumination deviation value, contrast compensation value, humidity-induced blur reduction amount, and sharpness restoration coefficient of each image region, the image data is adjusted to determine the standardized image data of the substation fittings.

3. The method for identifying corrosion defects in substation fittings as described in claim 2, characterized in that, The process of adjusting the image data based on the illumination deviation value, contrast compensation value, humidity-induced blur reduction amount, and sharpness restoration coefficient for each image region to determine the standardized image data of the substation fittings includes: Based on the illumination deviation value of each image region, the brightness of each image region is corrected to determine the corrected brightness value of each image region; Based on the contrast compensation value, humidity-induced blur reduction amount, and sharpness restoration coefficient of each image region, the sharpness and contrast of each image region are jointly optimized to determine the comprehensive optimized quality value of each image region. Based on the corrected brightness value and comprehensive optimized quality value of each image region, the pixel coordinates of the image data are arranged to determine the standardized image data of the substation fittings.

4. The method for identifying corrosion defects in substation fittings as described in claim 1, characterized in that, The step of extracting visual features of the substation fittings from the standardized image data and determining the target type and installation location of the substation fittings based on the visual features includes: The standardized image data is input into a pre-trained convolutional neural network to extract the edge features, texture features, and color distribution features of the substation hardware in order to determine the visual features of the substation hardware. Based on a preset support vector machine algorithm, the visual features of the substation fittings are classified to determine the target type and installation location of the substation fittings.

5. The method for identifying corrosion defects in substation fittings as described in claim 2, characterized in that, Based on the target type, installation location, and several environmental data of the substation fittings, a node retrieval is performed on a preset substation fitting knowledge graph to determine the environmental corrosion-related knowledge groups of the substation fittings, including: Based on the target type and installation location of the substation fittings, the node layer of the preset substation fitting knowledge graph is retrieved to obtain the substation fitting target type node corresponding to the target type and the substation fitting installation location node corresponding to the installation location. Based on the target type node and the installation location node of the substation fitting, the node association relationship of the relationship layer of the substation fitting knowledge graph is retrieved to determine the type association node set corresponding to the target type and the location association node set corresponding to the installation location, and the initial knowledge node group of the substation fitting is determined based on the type association node set and the location association node set. Based on a preset matching degree function, the matching degree between each type of environmental data and each environmental node in the node layer of the substation hardware knowledge graph is calculated; the environmental nodes are filtered based on the matching degree to determine the first environmental node corresponding to each type of environmental data; Based on the first environmental node, an environmental knowledge node group for the substation hardware is constructed; and the relationship layer of the substation hardware knowledge graph is retrieved to determine the corrosion association weight of each first environmental node. Based on the corrosion association weight of each first environmental node, and combined with the preset environmental data influence coefficient of each type of environmental data, the node association strength value between any associated node in the initial knowledge node group and any first environmental node in the environmental knowledge node group is determined. Based on the node association strength value, the initial knowledge node group and the environmental knowledge node group are filtered to determine the environmental corrosion association knowledge group of the substation hardware.

6. The method for identifying corrosion defects in substation fittings as described in claim 5, characterized in that, Based on the environmental corrosion association knowledge group, several corrosion influence paths of the substation fittings are constructed, and several key corrosion knowledge nodes of the substation fitting knowledge graph are extracted from the corrosion influence paths. Then, based on the key corrosion knowledge nodes, the corrosion association knowledge of the substation fittings is constructed, including: Starting from the first environmental node and ending at the substation hardware target type node, a path search is performed on the substation hardware knowledge graph based on a preset path search algorithm and preset path constraints to construct several corrosion impact paths of the substation hardware. Based on the relationship layer of the knowledge graph of the substation fittings, the path weight of each corrosion impact path is determined; based on the environmental data corresponding to the first environmental node of each corrosion impact path, the environmental data deviation rate of each corrosion impact path is determined. Based on the environmental data deviation rate and path weight of each corrosion impact path, and combined with the environmental data impact coefficient and corrosion association weight, the corrosion association strength of each corrosion impact path is determined. Based on the number of corrosion impact paths and the corrosion correlation strength of each corrosion impact path, the corrosion impact paths are screened to obtain several key corrosion impact paths. The frequency of each knowledge node in the key corrosion impact path is counted, and a number of key corrosion knowledge nodes are selected from the key corrosion impact path based on a preset frequency threshold. Based on each key corrosion impact path and the corrosion correlation strength of the key corrosion impact path, and combined with the key corrosion knowledge nodes, a corrosion correlation network is constructed; the corrosion correlation network is inferred based on the preset knowledge graph reasoning rules to construct the corrosion correlation knowledge of the substation fittings.

7. The method for identifying corrosion defects in substation fittings as described in claim 1, characterized in that, Based on the visual features and corrosion correlation knowledge, path identification is performed on the preset power defect identification map model to determine the effective path group of the substation hardware. Then, combining the visual features and corrosion correlation knowledge, the defects of the substation hardware are confirmed to obtain corrosion defect identification results, including: Based on the target type and installation location of the substation fittings, the nodes of the preset power defect identification graph model are matched with labels, and several graph model nodes are selected to form a graph model node subgroup. Construct a matching degree matrix between the visual features and the graph model node subgroup, and based on the matching degree matrix and the power defect identification graph model, filter the graph model nodes in the graph model node subgroup to obtain the candidate graph model node group; Based on the corrosion association knowledge and node subgroups, the attribute information of the graph model edges in the power defect identification graph model is analyzed, and several graph model edges are selected to form graph model edge subgroups; and based on the corrosion association knowledge, each graph model edge in the graph model edge subgroup is subjected to association strengthening processing to obtain the graph model edge attribute matrix. Based on the candidate graph model node group and graph model edge attribute matrix, the power defect identification graph model is used to identify the effective path group of the substation hardware. Based on the effective path group, visual features, and corrosion association knowledge of the substation fittings, the defects of the substation fittings are confirmed, and the corrosion defect identification results are obtained.

8. The method for identifying corrosion defects in substation fittings as described in claim 7, characterized in that, The process involves constructing a matching degree matrix between the visual features and the graph model node subgroup, and based on the matching degree matrix and the power defect identification graph model, filtering the graph model nodes in the graph model node subgroup to obtain a candidate graph model node group, including: The visual features are decomposed to construct several feature sub-vectors; Calculate the cosine similarity between each feature sub-vector and the feature attribute vector of the graph model node; calculate the ratio between each feature sub-vector and the feature attribute vector of the graph model node, and determine the attribute matching degree between each feature sub-vector and the graph model node; Based on the cosine similarity and attribute matching degree, the comprehensive matching degree between each feature sub-vector and the graph model node is determined, and then a matching degree matrix between the visual feature and the graph model node subgroup is constructed based on the comprehensive matching degree. The number of edges in the graph model of the power defect identification graph model is obtained, and the degree centrality of each graph model node in the graph model node subgroup is determined by combining the number of incoming edges and outgoing edges of each graph model node in the graph model node subgroup. Based on the matching degree matrix, determine the maximum matching degree of each graph model node in the graph model node subgroup; The degree centrality and maximum matching degree of each graph model node in the graph model node subgroup are weighted and summed to determine the comprehensive index of each graph model node in the graph model node subgroup; Based on the comprehensive index, the graph model nodes in the graph model node subgroup are filtered to obtain the candidate graph model node group.

9. The method for identifying corrosion defects in substation fittings as described in claim 7, characterized in that, The defect confirmation of the substation hardware based on the effective path group, visual features, and corrosion association knowledge yields corrosion defect identification results, including: Based on the path defect type mapping relationship library in the power defect identification graph model, each valid path in the valid path group of the substation fitting is judged to determine the defect type of each valid path; Based on the defect type of each effective path, the occurrence frequency of each defect type is determined, and the defect types are filtered based on the occurrence frequency to determine a candidate defect type group; Based on the corrosion association knowledge, calculate the degree of fit between each defect type in the candidate defect type group and the corrosion association knowledge, and determine the modified confidence level of each defect type based on the degree of fit; Based on the modified confidence level, the defect types are filtered to determine the confidence level defect type group; Based on the power defect identification map model, obtain the visual feature group corresponding to each confidence defect type in the confidence defect type group; Calculate the consistency score between the visual feature group corresponding to each confidence defect type in the confidence defect type group and the visual feature, and filter the confidence defect type group based on the consistency score to obtain the defect type group to be confirmed; Based on the power defect identification graph model, the path strength of each defect type to be identified in the defect type group is obtained, and the confidence assessment value of each defect type to be identified is calculated by combining the corrected confidence of each defect type to be identified. The type of defect to be confirmed with the highest confidence assessment value is taken as the corrosion defect type of the substation fitting, and the corrosion defect identification result is obtained by combining the visual features, corrosion association knowledge and effective path.

10. A system for identifying corrosion defects in substation fittings, characterized in that, include: The system includes an image standardization module, a visual feature extraction module, a knowledge graph retrieval module, a corrosion association knowledge construction module, and a corrosion defect identification result acquisition module. The image standardization module is used to acquire image data of substation fittings and several types of environmental data, and adjust the image data based on the environmental data to determine the standardized image data of the substation fittings. The visual feature extraction module is used to extract the visual features of the substation fittings from the standardized image data, and determine the target type and installation location of the substation fittings based on the visual features; The knowledge graph retrieval module is used to perform node retrieval on a preset substation hardware knowledge graph based on the target type, installation location, and several types of environmental data of the substation hardware, and to determine the environmental corrosion-related knowledge groups of the substation hardware. The corrosion association knowledge construction module is used to construct several corrosion influence paths of the substation fittings based on the environmental corrosion association knowledge group, and extract several key corrosion knowledge nodes of the substation fitting knowledge graph from the corrosion influence paths, and then construct the corrosion association knowledge of the substation fittings based on the key corrosion knowledge nodes. The corrosion defect identification result acquisition module is used to perform path identification on the preset power defect identification map model based on the visual features and the corrosion association knowledge, determine the effective path group of the substation hardware, and confirm the defect of the substation hardware by combining the visual features and corrosion association knowledge, so as to obtain the corrosion defect identification result.

Images